search

charlesq34 / pointnet

PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation
Other
4.73k stars 1.45k forks source link

T-nets #81

Closed ghost closed 6 years ago

ghost commented 6 years ago

Hi,

I was just wondering. Is there any particular reason for not enforcing orthogonality to the transform_XYZ net, but to do so in the transform_feat one? It seems to me that the first T-net can roam free to transform the input cloud, and only the second T-net contributes to the loss by orthogonal enforcement. Why not enforce the first T-net too?

Regards, Raul

charlesq34 commented 6 years ago

Hi @rdiazgarHP

You can also restrict the matrix to be rotations only. Currently the 3x3 matrix has a more free form such that it can also skew or scale the point clouds. For the 64x64 transformation matrix we found a regularization of orthogonality makes optimization more controllable and leads to better performance.

Best, Charles

ghost commented 6 years ago

I see. Thanks for the clarification!

adosar commented 7 months ago

Hi @rdiazgarHP

You can also restrict the matrix to be rotations only. Currently the 3x3 matrix has a more free form such that it can also skew or scale the point clouds. For the 64x64 transformation matrix we found a regularization of orthogonality makes optimization more controllable and leads to better performance.

Best, Charles

Hello @charlesq34! How can the rotation matrix be forced to represent only rotations? Moreover, would be a good idea to substitute T-Net with data augmentation (e.g. every kind of rotation)?